Data drift for LLM systems occurs when production inputs, retrieved content, user behavior, or external knowledge changes in ways that make existing prompts, models, or evals less reliable. The model may be the same, but the world around it has moved.
In agents and RAG systems, drift often appears in traces: new intents, new document types, changed APIs, shifted user language, or different tool failure patterns. Online evals and dataset refreshes help catch it.